“The 'Scientist AI' framework could be used as a foundation to design inherently safe AI agents, not just as an external guardrail.”